This document summarizes a study that uses a neural network called a Neocognitron to control the voltage of a synchronous generator. The Neocognitron was trained using data from simulating the generator with a conventional ST1 control. The neural network was then able to regulate the generator's voltage with better response time and robustness to large disturbances compared to the conventional control. Key aspects of the Neocognitron structure and learning algorithm are described. Simulation results show the neural control maintained voltage closer to the reference value than the ST1 control during a short circuit event.

1. The International Journal of Engineering
And Science (IJES)
||Volume|| 1 ||Issue|| 2 ||Pages|| 1-8 ||2012||
ISSN: 2319 – 1813 ISBN: 2319 – 1805
 Neural Control of Voltage for a Synchronous Generator
1,
Yoram Astudillo-Baza, 2,Manuel Alejandro López-Zepeda, 3,Sergio Baruch
Barragán-Gómez
1,2,3,
Instituto Politécnico Nacional
Escuela Superior de Ingeniería Mecánica y Eléctrica
Departamento de Ingeniería Eléctrica
Av. Instituto Politécnico Nacional s/n, Unidad Profesional “Adolfo López Mateos”
Edif. 2, Col. Lindavista, Del. Gustavo A. Madero, D.F. C.P. 07738
-------------------------------------------------Abstract------------------------------------------------------------
This article presents the regulation of the voltage of a synchronous generator using a neural network called
“Neocognitron”. A system of test machine bus -infinite is used, where the synchronous generator is connected to
an infinite bus through external impedance. The methodology is described used for the neural control, the
obtained results of this control type are presented and they are compared with results obtained with a
conventional control of the type ST1. It is shown that t he neural control has a better answer in the time and
bigger robustness before big perturbations.
Keywords – Cell, Conventional Control ST1, Neocognitron, Neural Control, Synchronous Generator.
-----------------------------------------------------------------------------------------------------------------------------------------------------
Date of Submission: 16, November, 2012 Date of Publication: 5,December 2012
-----------------------------------------------------------------------------------------------------------------------------------------------------
1. INTRODUCTION
The connection of the synchronous machines in parallel, demands that the generators have agreement
of phases, the same frequency and the same voltage in terminals. To be able to fulfil with these mentioned
requirements, equipment is needed that controls the speed, power and the voltage in terminals of the generators.
During the last twenty years, a great effort has been made to get new developments and technical of non
conventional control that can increase or to replace the conventional control techniques. A number of technical
of non conventional control has evolved, offering solutions to many problems of the control in the industry. This
is the essence of what the practical control has been called [1], that is a collection of techniques that the
engineers practicing has found efficiently and easily to use in the field The intelligent controls are characterized
by their ability of establishing a functional relationship between their inputs and outputs with empiric data,
without the resource of explicit models of the controlled systems. On the contrary of the conventional controls,
the intelligent controls can learn to remember and to make decisions.The intelligent controls can train to operate
indeed under conditions of uncertainty, they can respond autonomously to unexpected situations, without the
intervention of the operator of the system.The artificial neural networks are very appropriate tools for th e
modelling and control. They have the capacity to learn for the adjustment of internal parameters through the
experience, they have good properties of robustness with regard to incomplete information and with noise,
offering an attractive approach for the calculation of processes in the field and for the modelling of plants [2].
The properties that make to the artificial neural networks particularly applicable to control applications they are
the following ones:
To be nonlinear by nature, they are specially prepared for the control of nonlinear plants, they are
directly applicable to the control multivariable and they are relatively tolerant to flaws due to their parallel
structure, besides that they face new situations and they have the ability to generalize The neural network
denominated “Neocognitron” was developed by Kunihiko Fukushima with Sei Miyake’s assistance and
Takayuki I. around the years 70 and 80. The neocognitron is a design of network multi-layers that consists of
connections on form of cascade of many layers of cells. It is a network with hierarchy with layers that correspond
to simple and complex cells, with a very spread and located conn ection model among the layers The neocognitron
none alone it is the typical neuronal network with hierarchy, it is also the longest and the most complicated
www.theijes.com The IJES Page 1

2. Neural Control of Voltage for a Synchronous
neural network developed [3]. The basic function of the neocognitron is to act like a network of recognition of
models for images. The main advantage of the neocognitron is the ability to correctly not ec ognize single
learning models if not also models with partial movement, rotation and another type of distortion.
2. THE NEOCOGNITRON
The neocognitron is a neural network with parallel hierarchy designed for the recognition of hand
written characters [3-5]. These network this inspired one in the biological visual pattern of Hubel and Wiesel.
They acquire the ability to recognize robust visual models through the learning.The neocognitron carries out an
extraction process of characteristic by means of successive filtrates of the entrance image with layers connected
in cascade. Each layer extracts the appropriate characteristics from the exit of another previous one and it forms a
compressed representation of those extracted characteristics. The classification is achieved by a firm extraction.
2.1 The structure of the neocognitron
The structure of the neocognitron is shown in the Fig. 1 and emerges of a hierarchy of extracting
features. An appropriate stage of the neocognitron is believed for each hierarchy stage to extract features. The
network however contains an additional stage, labeled as stage 0 that is not used, in contrast with the stages but
high, for the extraction of features. The total number of stages of the neocognitron depends on the complexity of
the problem.
Fig. 1 The structure of the neocognitron.
In the stage 0 of the neocognitron alone the presented pattern is retained, also the alone stage 0
contains the input layer and it is proposed to retain the pattern presented to the network. In the highest stages in
the neocognitron the extraction of features is developed. The highest stages consist of two types of layers: the
S-layer composed by S-cells proposed for the extraction of features. The C-layer composed by C-cells, proposed
to assure a tolerance of changes of the features extracted in the S-layer previous. The cells denote the basic
element of operation. The cell is an element or neuron that it processes input values according to a certain rule
and it produces output values. There are three types of elementary cells that differ for their function: The
receptor cell is the simplest type of cell. Their function is alone to retain the presented models, the S-cell is the
most important type in the neocognitron and they has the function of extracting features of the input layer and
the C-cells whose function is to assure a tolerance of change in the features extracted by the S-cells.
2.2 Mathematical description of a S-cell.
The function of a S-cell is to extract a feature of a certain position in the input layer. And their
mathematical description is shown in the equation (1).
 K cl 1

1    al (v, k , K )  ucl 1 (n  v, K ) 
usl ( n, k )  rl     1
k 1 vAl
(1)
 r
1  l  bl ( k )  uvl ( n) 
 1  rl 
 
Where:
www.theijes.com The IJES Page 2

3. Neural Control of Voltage for a Synchronous
l Serial number of the stage.
n Position of the cell.
k Serial number of the cell plane.
rl Parameter of selectivity.
K cl 1 Number of cell planes in the C-layer previous.
v Position in the connection area.
Al Area of connection of the S-cell.
al , bl Weighted connections a y b≥0.
x si x  0
 x  
0 si x  0
2.3 Mathematical description of a C-cell.
The mathematical description of a C-cell is shown in the equation (2); each C-cell evaluates outputs of
the S-cells of a certain connection area.
 Kcl 1 
ucl (n, k )     jl ( K , k )   d l (v )  u sl (n  v, K )  (2)
 k 1 vDl 
Where:
l Serial number of the stage.
n Position of the cell.
k Serial number of the cell plane.
K sl 1 Number of cell planes in the S-layer previous.
jl Connection factor.
v Position in the connection area.
Dl
Area of connection of the C-cell.
dl
Weighted connections d ≥0.
 x x
 x  
si x  0
 x 
1   x , 0 si x  0
2.4. Learning of the neocognitron.
The learning is the process during which the net neuronal adapts to solve a desire task. The learning of
the network is controlled by a teacher or without supervisio n. This task is to determine that features will be
extracted in the stages of the network and to prepare the models of training before beginning the learning.
The learning of the neocognitron benefits stage for stage from the lowest stage in the network ad justing the
amendable weighted connections weighted a, weighted b according to the learned answer of the network to
represent a model of training.
The mathematical description of the learning is shown in the equation (3).
ˆ
al (v, K , k )  ql  cl (v)  ucl 1 (n  v, K )
ˆ
(3)
ˆ
b (k )  q  u (n) ˆ
l l vl
Where:
l Serial number of the stage.
al , bl
y cl Weighted connections a, b y c≥0.
ql
Learning coefficient.
ˆ
k Serial number of the cell plane.
ˆ
n Position of the activation cell.
www.theijes.com The IJES Page 3

4. Neural Control of Voltage for a Synchronous
3. THE NEOCOGNITRON TO REGULATE THE VOLTAGE OF THE SYNCHRONOUS GENERATOR
It is known that the function of an automatic regulator of voltage it is to adjust the level of excitement of
a synchronous machine continually before different conditions of operation of the machine with the purpose of
maintaining the voltage in terminals with the smallest quantity in possible variations, guaranteeing the stability of
the system under normal conditions of operation. Leaving of this objective, you proceed to apply the neural
methodology to regulate the voltage of the synchronous generator.
The proposed control strategy should make that the neocognitron learns the behavior of an automatic regulator
of voltage and to be able to improve this behavior it is necessary that the learning of the network is for different
points of operation of the system. The inputs of the neocognitron should be chosen with much care in such a
way that they generate the necessary minimum quantity of conditions to represent the dynamics of the system.
It is important to mention that several structures of the neocognitron exist and it stops effects of the simulation
the structure it was used shown in the figure 3. The structure of the proposed neocognitron consis ts of a input
layer of cells, two layers of S-cells and two layers of C-cells; one of these C-layers represents the layer of output
cells.The input layer as it is shown in the Fig. 2 this compound one for two input cells, one represents the error
defined by the difference between the reference voltage and the voltage in terminals of the generator and the
other one to the variation of the error defined by the difference between the current error and the state previous
of this error. The two layers of S-cells that are composed by 12 S-cells each one, and the two layers of C-cells that
a this compound one for 12 C-cells and the other layer of C-cells that represents to the output layer composed by
a C-cell that like leave in the Fig. 2 it contains the output variable represented by the field voltage, since this
variable allows us to generate the appropriate control law for our control system.
In the beginning of the learning with teacher one has a group of weighted a, weighted b in the network. Finally
the teacher adjusts the weighted of the cells according to the equations mentioned in the mathematical
description of learning. In principle, an input pattern is presented in the input layer and the fact spreads through
the network. Then it is allowed to increase the weighted to make adjustments according to the specific algorithm
of optimization.The input data for the training are an important part of the process of training because the group
of data is the only information that the neural network has to learn the task. The fact of training should embrace
the operation range for the neural network so that it estimates the wanted output.
Fig. 2 The structure of the neocognitron used for voltage regulation.
To obtain the data of training the synchronous machine it is simulated with a conventional control ST1,
generating input-output data for different operation conditions.To carry out the learning process we take the
algorithm of training of the neocognitron, providing as vectors or input patterns to the network the error and the
variation of the error, the error is given by the difference of the reference voltage and the real voltage obtained in
the terminals of the machine, the variation of the error is determined by the difference between the current error
and the previous error, therefore the input and output of the neural network for the learning process are shown in
the equation (4).
www.theijes.com The IJES Page 4

5. Neural Control of Voltage for a Synchronous
e(t )  Vref  VT (t )
e(t )  e(t )  e(t  1) (4)
u (t )  u (t )  u (t  1)
With these input vectors it is carried out we was carried out the learning process until reaching a
quadratic sum of the maximum error of 0.001 that allows to evaluate that the network has learned the behavior of
the system. The wanted output vector is more the variation of voltage of necessary field to maintain the voltage
in terminals him near the reference value similar to 1.0 for unit for our study.
The Fig. 3 shown the outline of the inverse control used for this simulation in two stages.
Stage a
Stage b
Fig. 3 Neurocontrol for the dynamic system; Stage a: Identification, Stage b: Con trol.
The inverse control offers a solution to the problem of control. The inverse control consists of two
stages: the first stage a; it is the identification stage where a neural network trains to behave that is to say as the
inverse of the system that will be controlled and the second stage b; is the control stage where the same network
is used to make the control of the system.
4. RES ULTS OF THE SIMULATION
For effects of the simulation the development of the control is based on a connected synchronous
generator to an infinite bus through a line of equivalent transmission, a mathematical model of 5º order is used for
the representation of the synchronous machine. Appendix A.The tests were carried out with the same parameters
of the synchronous machine and under the same operation conditions, allowing to make the comparison of the
system of neural regulation with the system of regulation of the type ST1 and this way to evaluate the index of
acting of each regulator. For the tests it is shown: the voltage in terminals Fig. 4, the angle of load of the rotor
Fig. 5, the active power Fig. 6, the voltage of field Fig. 7, and the index of acting of the voltage in terminals Fig.
8.For the application of the regulation systems we take several groups of conditions initials, they were proven for
damping of different times of duration. The time in that it happens the short circuit is of 1 second taking the bus
voltage to 0 for unit, for a damping with 12 duration cycles, the final time of the simulation is of 8 seconds. The
point of initial operation of the voltage in terminals that it is used for this test is Vt=1∟7° for unit and with an
active power of P = 0.8240 for unit.
www.theijes.com The IJES Page 5

6. Neural Control of Voltage for a Synchronous
Fig. 4 Voltage in terminals.
Fig. 5 Angle of load of the rotor.
Fig. 6 Active Power.
Fig. 7 Field Voltage.
Fig. 8 Index of the quadratic error of the voltage in terminals .
www.theijes.com The IJES Page 6

7. Neural Control of Voltage for a Synchronous
5. CONCLUS ION
The motivation of this work is to take advantage of the modern technology of the intelligent control to
obtain a better potential in the stability of the system of power low short circuits.The proposed regulator enjoys
the general advantages of the neural networks like generalization capacity, tolerance to damping and the property
of adaptability due to its learning The results of the simulation of the regulator proposed on several operation
conditions and disturbances, demonstrated that the syst em of voltage regulation based on the neocognitron
obtains a better answer in the time that the conventional regulator ST1, as well as a bigger robustnessWe can
conclude that the voltage in terminals when the control neuronal is used it is more constant an d with smaller
oscillations that when the conventional control is used.
APPENDIX A
The differential equations of the pattern of 5º order, used in this work are [6].
p r   0 s
Mps  Kd s  Tm  Te
   
Td 0 peq  V f  ( X d  X d )id  eq
Td0 peq  eq  ( X d  X d )id  eq
     
Tq0 ped  ( X q  X q )iq  ed
   
 
ed  Vd  raid  X qiq
 
eq  Vq  raiq  X d id
   
Te  ed id  eq iq  ( X d  X q )id iq
Without loss of generality, the transmission line is described by an equivalent impedance of Thevenin.
Therefore, the terminal voltage and their components in direct axis and axis in quadrature, they are the following
ones [6].
Vd  V Sen r  reid  X eiq
Vq  VCos r  reiq  X eid
2 2 2
VT  Vd  Vq
The data of the system machine - infinite bus is shown in the Table 1. For the regulator of voltage ST1, these
they are taken of the reference [6], likewise the data of the transmission line corresponds to a gen erator of 645
MVA in for unit [6].
Table 1. Data of the Low System Prove
0 377 Xe 0.15
M 5.5294 re 0.01
Kd 3.0 KA 400.0
T’do 5.66 TA 0.02
T’’do 0.041 KF 0.008
T’’qo 0.065 TF 1.0
Xd 1.904 TX 0.025
X’d 0.312 P 0.824
X’’d 0.266 Xq 1.881
X’’q 0.260 Ra 0.0
www.theijes.com The IJES Page 7

8. Neural Control of Voltage for a Synchronous
REFERENCES
[1] R. E. King. Computational Intelligence in Control Engineering. (University of Patras, Greece, Marcel
Dekker, Inc. New York, 1999).
[2] M. Norgaard, O. Ravn, N. K. Poulsen and L. K. Hansen. Neural Networks for Modelling and Control of
Dynamic Systems. (Springer, 2000).
[3] K. Fukushima: “Neocognitron: A hierarchical neural network capable of visual pattern recognition”, Neural
Networks, 1[2], pp. 119-130 (1988).
[4] K. Fukushima y S. Miyake: “Neocognitron: A new algorithm pattern recognition tolerant of deformations
and shifts in position”, Pattern Recognition, 15[6], pp. 445-469 (1982).
[5] K. Fukushima, “Neural network model for a mechanism of pattern recognition unaffected by shift in
position”, Trans. IECE Japan, Vol. 62-A, no 10, pp. 658-665 (1979).
[6] D. Xia., G.T. Heydt. Self-Tuning Controller For Generator Excitation Control. IEEE Transaction on
Power Apparatus and Systems. Vol. PAS-102. No.6 June 1983 pp. 1877-1884.
Biographies
Yoram Astudillo-Baza. He received Master in Electrical Engineering from ESIME-IPN, Mexico in 2005. He is
currently a Mathematics professor at Department of Electrical Engineering of ESIME-IPN. His research interests
Analysis and Control of Electrical Power Systems, Electrical Machines, Intelligen t Control, Power Generation,
Cogeneration.
Manuel Alejandro López-Zepeda. He received BsC and Masters in Electrical Engineering from ESIME-IPN,
Mexico in 2002 and 2006. He’s currently a Computer Science professor at ESIME-IPN. His research interests span
state estimation in electric power systems, intelligent control and neuronal networks.
Sergio Baruch Barragán-Gómez. He received Masters in Electrical Engineering from SEPI-ESIME- IPN, Mexico
in 2004. He is currently a electric power systems professor at Department of Electrical Engineering of ESIME-IPN.
His research interests: open software, analysis and optimization of electrical power systems.
www.theijes.com The IJES Page 8